Block unveils Dessa, a voice AI that learns only by listening to humans talk
A new AI from Block talks and listens at the same time — and it was trained only on human speech, not on text.
How it works
A voice AI that listens and speaks at the same time, trained only on human conversation audio, not text.
- You speak naturally to the AI
- It listens while you are still talking
- It jumps in, laughs, or reacts in real time
- It runs right on your laptop locally
What you could do with it
- A developer could run full-duplex voice conversations locally on a MacBook Pro with an M4 Max chip, at speeds faster than real time 1.
- A researcher could explore how a speech-only model learns new languages and performs in-context learning, since Block's project page invites questions about both 2.
- A user could have a realistic, emotional back-and-forth conversation with an AI that listens and speaks at the same time, using cues like laughter, sighs, and interruptions 2.
Guess
Can Dessa listen and speak at the same time?
Dessa can listen and speak simultaneously, like a real person interrupting or nodding along.
An exampleInvented to explain it
Maya practicing a conversation alone at home
MayaHey, did you hear that crazy news today?
DessaNo wait, tell me — what happened?
MayaMy friend just got accepted into that program!
DessaOh my god, that is amazing!
MayaRight? I am so happy for them.
DessaYeah totally! That is awesome news.
Maya left feeling like she had a natural back-and-forth conversation with an AI.
How Dessa is different
Most voice AI systems today work in three stages: they turn your voice into text, run the text through a language model, then turn the answer back into speech 1. Block says that chain throws away emotion, sarcasm, laughter and sighs — the things that make a conversation feel human 2.
Dessa takes a different path. It goes straight from sound to sound. Block trained it from scratch on nothing but recordings of people talking to each other 2. No text. No pretrained models. "We believe voice is the most natural way to interact with intelligence," the company wrote on the project page .
Dessa is full-duplex, which means it can listen and speak at the same time. Real people do this all the time — they nod along with a quick "yeah" while someone else is still talking, or they jump in before the other person finishes a sentence. Most current voice AIs can't do that; they wait for you to stop, then respond 3. Block says Dessa learned when to cut in and when to backchannel just by hearing real conversations 12.
What the demos show
Block shared several live recordings of people talking to Dessa. In one, it laughs mid-sentence — something the company says other models only do on cue after you finish speaking 2. In another, it gasps "oh my god" when surprised and sounds tearfully grateful . When asked to speak up, it gets louder . One clip even shows Dessa talking to OpenAI's GPT-Live-1 1. Nobody guided any of the clips, Block says .
How efficient is it?
1 According to Block, Dessa acquired language using hundreds of times fewer words than what large text-based AI models require. The company trained it on a single NVIDIA GB200 NVL72 rack — 72 Blackwell GPUs linked with 36 Grace CPUs — far fewer chips than rival full-duplex models . For comparison, Kyutai's Moshi, another open full-duplex model, used 1,016 H100 GPUs and 7 million hours of pretraining audio 3.
Dessa can run locally on a MacBook Pro with an M4 Max chip or better, at speeds faster than real time 1.
What we still don't know
Despite the demos, Block has not shared several key details: no benchmark scores, no parameter count, no model weights, no research paper, no API and no price 1. Access is currently a waitlist-only research preview . The model was announced on October 23, 2026 .
How good is it
Block has published zero benchmark scores, no parameter count, and no model weights for Dessa, so there are no independent test results to evaluate 1. The company shared live demo recordings showing natural laughter, mid-sentence interruptions, and expressive reactions, but these were not benchmarked against rival models . According to Block, Dessa learned language using hundreds of times fewer words than large text-based AI models require, though this claim has not been independently verified . Compared rivals like Moshi and GPT-Live-1 have published latency numbers, but Dessa's performance metrics remain undisclosed .
Key facts
- Dessa is a full-duplex speech-to-speech AI trained only on human conversation audio
- Block trained it on 1 NVIDIA GB200 NVL72 rack (72 GPUs); rival Moshi used 1,016 H100s
- Runs locally on a MacBook Pro M4 Max or better, faster than real time
- No benchmark scores, weights, paper, API or price have been published
- Access is a waitlist-only research preview announced October 23, 2026
Sources
Keep reading
The weekThis Week in AI New Models Flood In and Agents Get SaferHerald of AI▶The weekThis Week in AI New Models Flood In and Agents Get Safer
Big open-weight releases, speed breakthroughs, and a push to make AI agents more reliable and secure.
Related stories
How it worksA new engine makes small AI models respond in under 50 millisecondsHerald of AI▶
A new engine makes small AI models respond in under 50 milliseconds
Liquid AI says its tool lets generative AI replace old models for fraud checks, payments and search — without slowing things down.
New modelsNVIDIA Kumo Relational beats supervised ML on relational data without trainingHerald of AI▶
NVIDIA Kumo Relational beats supervised ML on relational data without training
A new pre-trained foundation model predicts outcomes directly on enterprise databases, requiring no feature engineering or task-specific training.
The weekThis Week in AI Agentic Tools, Cheaper Models, and Growing PainsHerald of AI▶
This Week in AI Agentic Tools, Cheaper Models, and Growing Pains
OpenAI reports agentic adoption in the EU, new models push one-million-token outputs, and developers wrestle with faster-but-opaque code.
How it worksDeepSeek Open-Sources DeepInfer, Its High-Performance Inference EngineHerald of AI▶
DeepSeek Open-Sources DeepInfer, Its High-Performance Inference Engine
The Chinese AI lab released the framework behind its production services under an Apache 2.0 license, giving developers access to large-scale serving optimizations.